Section 21.1.5. Preventing errors

(1) Try to obtain the best possible set of data and the best possible set of phases for those data. If the structure has noncrystallographic symmetry (or if multiple crystal forms are available), use electron-density averaging to remove model bias and to reduce phase errors (Kleywegt & Read, 1997). In the absence of noncrystallographic symmetry, use maps that are biased by the model as little as possible [e.g. σ_A-weighted (Read, 1986) or omit maps (Bhat & Cohen, 1984; Bhat, 1988; Hodel et al., 1992)]. If experimental phase information is available, keep and consult the experimental map(s). Experimental phases can also be used throughout the refinement process to alleviate or prevent some problems.

(2) Use databases to construct the initial model (or new parts of the model; Jones et al., 1991; Kleywegt & Jones, 1998). All the crystallographer needs to do is to roughly place the C^α atoms in the density. The model-building program can then `recycle' well refined high-resolution structures to place the main-chain atoms. Similarly, side-chain conformations should initially be chosen from the set of preferred rotamers for each residue type, perhaps in combination with a rigid-body rotation of the entire residue around its C^α atom and/or with minor adjustment of the torsion angles of long side chains (arginine, lysine etc.).

(3) After every cycle of refinement, carry out a critical analysis of the quality of the current model. This entails the calculation of properties such as those discussed in Section 21.1.3 and the inspection of the residues that are outliers for any of them, as described in Section 21.1.4. Be conservative during rebuilding, especially when the model is incomplete and possibly full of errors.

(4) Design a refinement protocol that is appropriate for the available data. If NCS restraints do not give a significantly better free R value than NCS constraints, then use constraints. If NCS restraints are to be employed, then use the experimental map to design a suitable NCS-restraint scheme (Kleywegt, 1999). Avoid the temptation to model alternative conformations in low-resolution maps or to place putative solvent molecules in every local maximum of the (F_o − F_c, α_c) difference map. In other words, be conservative and remember that the maxim `where freedom is given, liberties are taken' is highly applicable to refinement programs (Hendrickson & Konnert, 1980; Kleywegt & Jones, 1995b).

(5) Adopt methodological advances as soon as they become available. Several innovations have only been slowly accepted by the mainstream (e.g. the use of databases in building and rebuilding, the use of the free R value, the use of electron-density averaging in molecular-replacement cases, bulk-solvent modelling). The most prominent recent development is the use of likelihood-based refinement programs (Bricogne & Irwin, 1996; Pannu & Read, 1996; Murshudov et al., 1997; Adams et al., 1997; Pannu et al., 1998). These programs produce better models and maps and considerably reduce over-fitting (as assessed by the difference between the free and conventional R values).

(6) Most importantly, the crystallographer should be hypercritical towards the fruits of his or her own labour. Every intermediate model is a hypothesis to be shot down (Jones & Kjeldgaard, 1994). The crystallographer should be more critical than the supervisor, the supervisor more critical than the referee and the referee more critical than the casual reader. It goes without saying that the reader, casual or not, should have access to model coordinates, experimental data and electron-density maps!